Learning Joint Interventional Effects from Single-Variable Interventions in Additive Models
Authors: Armin Kekić, Sergio Hernan Garrido Mejia, Bernhard Schölkopf
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on synthetic data demonstrate that our method achieves performance comparable to models trained directly on joint interventional data, outperforming a purely observational estimator. We compare three approaches: (i) Our Intervention Generalization method, training the estimator (17) on observational and single-intervention data (Section 6). (ii) An estimator trained directly on joint interventional data (topline). (iii) A naive estimator trained on pooled dataset of all observational and single-interventional data. (iv) An estimator trained solely on observational data. |
| Researcher Affiliation | Collaboration | 1Max Planck Institute for Intelligent Systems, T ubingen, Germany 2Amazon Research, T ubingen, Germany 3T ubingen AI Center, T ubingen, Germany 4ELLIS Institute, T ubingen, Germany. |
| Pseudocode | No | The paper describes the methodology in prose and mathematical notation within the main text and appendix sections, without presenting any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The code used for the experiments is available at github.com/akekic/intervention-generalization. |
| Open Datasets | No | Synthetic Data-Generating Process. We sample a structural causal model with five actions and confounders and causal relationships as shown in Figure 3. The structural assignments are second order polynomials with randomly sampled coefficients. The exogenous noises are Gaussian, uniform or logistic. The corresponding parameters are drawn at random before each experiment run. |
| Dataset Splits | Yes | We split each dataset into 80% training and 20% test data. |
| Hardware Specification | No | The paper does not specify any particular hardware components such as GPU models, CPU types, or memory used for running the experiments. It only mentions the number of data points for various experimental runs. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in the implementation of the experiments. |
| Experiment Setup | Yes | We train third order polynomial estimator functions (17) as outlined in Section 6. We regularize the estimators using Ridge regression and find the optimal regularization parameter through 3-fold cross validation for each estimator. In order to satisfy Assumption 1 on the support of the interventions, we sample single interventions and joint interventions on the action variables to match the observational distributions. That is, we sample intervention values from a normal distribution: Aint k ~ N(µˆk, σˆ2 k) where µˆk and σˆk are the empirical mean and standard deviation of Ak in the observational data, respectively. For joint interventions, the intervention values are sampled independently for each action variable, following the same distribution as in the single intervention case. |